src/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecCoul_VdwNone_GeomW3W3_avx_128_fma_double.c

   1 /*
   2  * Note: this file was generated by the Gromacs avx_128_fma_double kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_avx_128_fma_double.h"
  34 #include "kernelutil_x86_avx_128_fma_double.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW3W3_VF_avx_128_fma_double
  38  * Electrostatics interaction: Coulomb
  39  * VdW interaction:            None
  40  * Geometry:                   Water3-Water3
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecCoul_VdwNone_GeomW3W3_VF_avx_128_fma_double
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB;
  61     int              j_coord_offsetA,j_coord_offsetB;
  62     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  63     real             rcutoff_scalar;
  64     real             *shiftvec,*fshift,*x,*f;
  65     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  66     int              vdwioffset0;
  67     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  68     int              vdwioffset1;
  69     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  70     int              vdwioffset2;
  71     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  72     int              vdwjidx0A,vdwjidx0B;
  73     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  74     int              vdwjidx1A,vdwjidx1B;
  75     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  76     int              vdwjidx2A,vdwjidx2B;
  77     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  78     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  79     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  80     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  81     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  82     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  83     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  84     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  85     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  86     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  87     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  88     real             *charge;
  89     __m128d          dummy_mask,cutoff_mask;
  90     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
  91     __m128d          one     = _mm_set1_pd(1.0);
  92     __m128d          two     = _mm_set1_pd(2.0);
  93     x                = xx[0];
  94     f                = ff[0];
  95
  96     nri              = nlist->nri;
  97     iinr             = nlist->iinr;
  98     jindex           = nlist->jindex;
  99     jjnr             = nlist->jjnr;
 100     shiftidx         = nlist->shift;
 101     gid              = nlist->gid;
 102     shiftvec         = fr->shift_vec[0];
 103     fshift           = fr->fshift[0];
 104     facel            = _mm_set1_pd(fr->epsfac);
 105     charge           = mdatoms->chargeA;
 106
 107     /* Setup water-specific parameters */
 108     inr              = nlist->iinr[0];
 109     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 110     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 111     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 112
 113     jq0              = _mm_set1_pd(charge[inr+0]);
 114     jq1              = _mm_set1_pd(charge[inr+1]);
 115     jq2              = _mm_set1_pd(charge[inr+2]);
 116     qq00             = _mm_mul_pd(iq0,jq0);
 117     qq01             = _mm_mul_pd(iq0,jq1);
 118     qq02             = _mm_mul_pd(iq0,jq2);
 119     qq10             = _mm_mul_pd(iq1,jq0);
 120     qq11             = _mm_mul_pd(iq1,jq1);
 121     qq12             = _mm_mul_pd(iq1,jq2);
 122     qq20             = _mm_mul_pd(iq2,jq0);
 123     qq21             = _mm_mul_pd(iq2,jq1);
 124     qq22             = _mm_mul_pd(iq2,jq2);
 125
 126     /* Avoid stupid compiler warnings */
 127     jnrA = jnrB = 0;
 128     j_coord_offsetA = 0;
 129     j_coord_offsetB = 0;
 130
 131     outeriter        = 0;
 132     inneriter        = 0;
 133
 134     /* Start outer loop over neighborlists */
 135     for(iidx=0; iidx<nri; iidx++)
 136     {
 137         /* Load shift vector for this list */
 138         i_shift_offset   = DIM*shiftidx[iidx];
 139
 140         /* Load limits for loop over neighbors */
 141         j_index_start    = jindex[iidx];
 142         j_index_end      = jindex[iidx+1];
 143
 144         /* Get outer coordinate index */
 145         inr              = iinr[iidx];
 146         i_coord_offset   = DIM*inr;
 147
 148         /* Load i particle coords and add shift vector */
 149         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 150                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 151
 152         fix0             = _mm_setzero_pd();
 153         fiy0             = _mm_setzero_pd();
 154         fiz0             = _mm_setzero_pd();
 155         fix1             = _mm_setzero_pd();
 156         fiy1             = _mm_setzero_pd();
 157         fiz1             = _mm_setzero_pd();
 158         fix2             = _mm_setzero_pd();
 159         fiy2             = _mm_setzero_pd();
 160         fiz2             = _mm_setzero_pd();
 161
 162         /* Reset potential sums */
 163         velecsum         = _mm_setzero_pd();
 164
 165         /* Start inner kernel loop */
 166         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 167         {
 168
 169             /* Get j neighbor index, and coordinate index */
 170             jnrA             = jjnr[jidx];
 171             jnrB             = jjnr[jidx+1];
 172             j_coord_offsetA  = DIM*jnrA;
 173             j_coord_offsetB  = DIM*jnrB;
 174
 175             /* load j atom coordinates */
 176             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 177                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 178
 179             /* Calculate displacement vector */
 180             dx00             = _mm_sub_pd(ix0,jx0);
 181             dy00             = _mm_sub_pd(iy0,jy0);
 182             dz00             = _mm_sub_pd(iz0,jz0);
 183             dx01             = _mm_sub_pd(ix0,jx1);
 184             dy01             = _mm_sub_pd(iy0,jy1);
 185             dz01             = _mm_sub_pd(iz0,jz1);
 186             dx02             = _mm_sub_pd(ix0,jx2);
 187             dy02             = _mm_sub_pd(iy0,jy2);
 188             dz02             = _mm_sub_pd(iz0,jz2);
 189             dx10             = _mm_sub_pd(ix1,jx0);
 190             dy10             = _mm_sub_pd(iy1,jy0);
 191             dz10             = _mm_sub_pd(iz1,jz0);
 192             dx11             = _mm_sub_pd(ix1,jx1);
 193             dy11             = _mm_sub_pd(iy1,jy1);
 194             dz11             = _mm_sub_pd(iz1,jz1);
 195             dx12             = _mm_sub_pd(ix1,jx2);
 196             dy12             = _mm_sub_pd(iy1,jy2);
 197             dz12             = _mm_sub_pd(iz1,jz2);
 198             dx20             = _mm_sub_pd(ix2,jx0);
 199             dy20             = _mm_sub_pd(iy2,jy0);
 200             dz20             = _mm_sub_pd(iz2,jz0);
 201             dx21             = _mm_sub_pd(ix2,jx1);
 202             dy21             = _mm_sub_pd(iy2,jy1);
 203             dz21             = _mm_sub_pd(iz2,jz1);
 204             dx22             = _mm_sub_pd(ix2,jx2);
 205             dy22             = _mm_sub_pd(iy2,jy2);
 206             dz22             = _mm_sub_pd(iz2,jz2);
 207
 208             /* Calculate squared distance and things based on it */
 209             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 210             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 211             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 212             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 213             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 214             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 215             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 216             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 217             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 218
 219             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 220             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 221             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 222             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 223             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 224             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 225             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 226             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 227             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 228
 229             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 230             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 231             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 232             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 233             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 234             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 235             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 236             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 237             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 238
 239             fjx0             = _mm_setzero_pd();
 240             fjy0             = _mm_setzero_pd();
 241             fjz0             = _mm_setzero_pd();
 242             fjx1             = _mm_setzero_pd();
 243             fjy1             = _mm_setzero_pd();
 244             fjz1             = _mm_setzero_pd();
 245             fjx2             = _mm_setzero_pd();
 246             fjy2             = _mm_setzero_pd();
 247             fjz2             = _mm_setzero_pd();
 248
 249             /**************************
 250              * CALCULATE INTERACTIONS *
 251              **************************/
 252
 253             /* COULOMB ELECTROSTATICS */
 254             velec            = _mm_mul_pd(qq00,rinv00);
 255             felec            = _mm_mul_pd(velec,rinvsq00);
 256
 257             /* Update potential sum for this i atom from the interaction with this j atom. */
 258             velecsum         = _mm_add_pd(velecsum,velec);
 259
 260             fscal            = felec;
 261
 262             /* Update vectorial force */
 263             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 264             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 265             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 266
 267             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 268             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 269             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 270
 271             /**************************
 272              * CALCULATE INTERACTIONS *
 273              **************************/
 274
 275             /* COULOMB ELECTROSTATICS */
 276             velec            = _mm_mul_pd(qq01,rinv01);
 277             felec            = _mm_mul_pd(velec,rinvsq01);
 278
 279             /* Update potential sum for this i atom from the interaction with this j atom. */
 280             velecsum         = _mm_add_pd(velecsum,velec);
 281
 282             fscal            = felec;
 283
 284             /* Update vectorial force */
 285             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 286             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 287             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 288
 289             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 290             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 291             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 292
 293             /**************************
 294              * CALCULATE INTERACTIONS *
 295              **************************/
 296
 297             /* COULOMB ELECTROSTATICS */
 298             velec            = _mm_mul_pd(qq02,rinv02);
 299             felec            = _mm_mul_pd(velec,rinvsq02);
 300
 301             /* Update potential sum for this i atom from the interaction with this j atom. */
 302             velecsum         = _mm_add_pd(velecsum,velec);
 303
 304             fscal            = felec;
 305
 306             /* Update vectorial force */
 307             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 308             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 309             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 310
 311             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 312             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 313             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 314
 315             /**************************
 316              * CALCULATE INTERACTIONS *
 317              **************************/
 318
 319             /* COULOMB ELECTROSTATICS */
 320             velec            = _mm_mul_pd(qq10,rinv10);
 321             felec            = _mm_mul_pd(velec,rinvsq10);
 322
 323             /* Update potential sum for this i atom from the interaction with this j atom. */
 324             velecsum         = _mm_add_pd(velecsum,velec);
 325
 326             fscal            = felec;
 327
 328             /* Update vectorial force */
 329             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 330             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 331             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 332
 333             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 334             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 335             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 336
 337             /**************************
 338              * CALCULATE INTERACTIONS *
 339              **************************/
 340
 341             /* COULOMB ELECTROSTATICS */
 342             velec            = _mm_mul_pd(qq11,rinv11);
 343             felec            = _mm_mul_pd(velec,rinvsq11);
 344
 345             /* Update potential sum for this i atom from the interaction with this j atom. */
 346             velecsum         = _mm_add_pd(velecsum,velec);
 347
 348             fscal            = felec;
 349
 350             /* Update vectorial force */
 351             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 352             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 353             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 354
 355             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 356             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 357             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 358
 359             /**************************
 360              * CALCULATE INTERACTIONS *
 361              **************************/
 362
 363             /* COULOMB ELECTROSTATICS */
 364             velec            = _mm_mul_pd(qq12,rinv12);
 365             felec            = _mm_mul_pd(velec,rinvsq12);
 366
 367             /* Update potential sum for this i atom from the interaction with this j atom. */
 368             velecsum         = _mm_add_pd(velecsum,velec);
 369
 370             fscal            = felec;
 371
 372             /* Update vectorial force */
 373             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 374             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 375             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 376
 377             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 378             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 379             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 380
 381             /**************************
 382              * CALCULATE INTERACTIONS *
 383              **************************/
 384
 385             /* COULOMB ELECTROSTATICS */
 386             velec            = _mm_mul_pd(qq20,rinv20);
 387             felec            = _mm_mul_pd(velec,rinvsq20);
 388
 389             /* Update potential sum for this i atom from the interaction with this j atom. */
 390             velecsum         = _mm_add_pd(velecsum,velec);
 391
 392             fscal            = felec;
 393
 394             /* Update vectorial force */
 395             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 396             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 397             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 398
 399             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 400             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 401             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 402
 403             /**************************
 404              * CALCULATE INTERACTIONS *
 405              **************************/
 406
 407             /* COULOMB ELECTROSTATICS */
 408             velec            = _mm_mul_pd(qq21,rinv21);
 409             felec            = _mm_mul_pd(velec,rinvsq21);
 410
 411             /* Update potential sum for this i atom from the interaction with this j atom. */
 412             velecsum         = _mm_add_pd(velecsum,velec);
 413
 414             fscal            = felec;
 415
 416             /* Update vectorial force */
 417             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 418             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 419             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 420
 421             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 422             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 423             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 424
 425             /**************************
 426              * CALCULATE INTERACTIONS *
 427              **************************/
 428
 429             /* COULOMB ELECTROSTATICS */
 430             velec            = _mm_mul_pd(qq22,rinv22);
 431             felec            = _mm_mul_pd(velec,rinvsq22);
 432
 433             /* Update potential sum for this i atom from the interaction with this j atom. */
 434             velecsum         = _mm_add_pd(velecsum,velec);
 435
 436             fscal            = felec;
 437
 438             /* Update vectorial force */
 439             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 440             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 441             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 442
 443             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 444             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 445             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 446
 447             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 448
 449             /* Inner loop uses 279 flops */
 450         }
 451
 452         if(jidx<j_index_end)
 453         {
 454
 455             jnrA             = jjnr[jidx];
 456             j_coord_offsetA  = DIM*jnrA;
 457
 458             /* load j atom coordinates */
 459             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 460                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 461
 462             /* Calculate displacement vector */
 463             dx00             = _mm_sub_pd(ix0,jx0);
 464             dy00             = _mm_sub_pd(iy0,jy0);
 465             dz00             = _mm_sub_pd(iz0,jz0);
 466             dx01             = _mm_sub_pd(ix0,jx1);
 467             dy01             = _mm_sub_pd(iy0,jy1);
 468             dz01             = _mm_sub_pd(iz0,jz1);
 469             dx02             = _mm_sub_pd(ix0,jx2);
 470             dy02             = _mm_sub_pd(iy0,jy2);
 471             dz02             = _mm_sub_pd(iz0,jz2);
 472             dx10             = _mm_sub_pd(ix1,jx0);
 473             dy10             = _mm_sub_pd(iy1,jy0);
 474             dz10             = _mm_sub_pd(iz1,jz0);
 475             dx11             = _mm_sub_pd(ix1,jx1);
 476             dy11             = _mm_sub_pd(iy1,jy1);
 477             dz11             = _mm_sub_pd(iz1,jz1);
 478             dx12             = _mm_sub_pd(ix1,jx2);
 479             dy12             = _mm_sub_pd(iy1,jy2);
 480             dz12             = _mm_sub_pd(iz1,jz2);
 481             dx20             = _mm_sub_pd(ix2,jx0);
 482             dy20             = _mm_sub_pd(iy2,jy0);
 483             dz20             = _mm_sub_pd(iz2,jz0);
 484             dx21             = _mm_sub_pd(ix2,jx1);
 485             dy21             = _mm_sub_pd(iy2,jy1);
 486             dz21             = _mm_sub_pd(iz2,jz1);
 487             dx22             = _mm_sub_pd(ix2,jx2);
 488             dy22             = _mm_sub_pd(iy2,jy2);
 489             dz22             = _mm_sub_pd(iz2,jz2);
 490
 491             /* Calculate squared distance and things based on it */
 492             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 493             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 494             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 495             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 496             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 497             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 498             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 499             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 500             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 501
 502             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 503             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 504             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 505             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 506             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 507             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 508             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 509             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 510             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 511
 512             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 513             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 514             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 515             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 516             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 517             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 518             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 519             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 520             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 521
 522             fjx0             = _mm_setzero_pd();
 523             fjy0             = _mm_setzero_pd();
 524             fjz0             = _mm_setzero_pd();
 525             fjx1             = _mm_setzero_pd();
 526             fjy1             = _mm_setzero_pd();
 527             fjz1             = _mm_setzero_pd();
 528             fjx2             = _mm_setzero_pd();
 529             fjy2             = _mm_setzero_pd();
 530             fjz2             = _mm_setzero_pd();
 531
 532             /**************************
 533              * CALCULATE INTERACTIONS *
 534              **************************/
 535
 536             /* COULOMB ELECTROSTATICS */
 537             velec            = _mm_mul_pd(qq00,rinv00);
 538             felec            = _mm_mul_pd(velec,rinvsq00);
 539
 540             /* Update potential sum for this i atom from the interaction with this j atom. */
 541             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 542             velecsum         = _mm_add_pd(velecsum,velec);
 543
 544             fscal            = felec;
 545
 546             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 547
 548             /* Update vectorial force */
 549             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 550             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 551             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 552
 553             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 554             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 555             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 556
 557             /**************************
 558              * CALCULATE INTERACTIONS *
 559              **************************/
 560
 561             /* COULOMB ELECTROSTATICS */
 562             velec            = _mm_mul_pd(qq01,rinv01);
 563             felec            = _mm_mul_pd(velec,rinvsq01);
 564
 565             /* Update potential sum for this i atom from the interaction with this j atom. */
 566             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 567             velecsum         = _mm_add_pd(velecsum,velec);
 568
 569             fscal            = felec;
 570
 571             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 572
 573             /* Update vectorial force */
 574             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 575             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 576             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 577
 578             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 579             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 580             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 581
 582             /**************************
 583              * CALCULATE INTERACTIONS *
 584              **************************/
 585
 586             /* COULOMB ELECTROSTATICS */
 587             velec            = _mm_mul_pd(qq02,rinv02);
 588             felec            = _mm_mul_pd(velec,rinvsq02);
 589
 590             /* Update potential sum for this i atom from the interaction with this j atom. */
 591             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 592             velecsum         = _mm_add_pd(velecsum,velec);
 593
 594             fscal            = felec;
 595
 596             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 597
 598             /* Update vectorial force */
 599             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 600             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 601             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 602
 603             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 604             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 605             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 606
 607             /**************************
 608              * CALCULATE INTERACTIONS *
 609              **************************/
 610
 611             /* COULOMB ELECTROSTATICS */
 612             velec            = _mm_mul_pd(qq10,rinv10);
 613             felec            = _mm_mul_pd(velec,rinvsq10);
 614
 615             /* Update potential sum for this i atom from the interaction with this j atom. */
 616             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 617             velecsum         = _mm_add_pd(velecsum,velec);
 618
 619             fscal            = felec;
 620
 621             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 622
 623             /* Update vectorial force */
 624             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 625             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 626             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 627
 628             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 629             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 630             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 631
 632             /**************************
 633              * CALCULATE INTERACTIONS *
 634              **************************/
 635
 636             /* COULOMB ELECTROSTATICS */
 637             velec            = _mm_mul_pd(qq11,rinv11);
 638             felec            = _mm_mul_pd(velec,rinvsq11);
 639
 640             /* Update potential sum for this i atom from the interaction with this j atom. */
 641             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 642             velecsum         = _mm_add_pd(velecsum,velec);
 643
 644             fscal            = felec;
 645
 646             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 647
 648             /* Update vectorial force */
 649             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 650             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 651             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 652
 653             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 654             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 655             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 656
 657             /**************************
 658              * CALCULATE INTERACTIONS *
 659              **************************/
 660
 661             /* COULOMB ELECTROSTATICS */
 662             velec            = _mm_mul_pd(qq12,rinv12);
 663             felec            = _mm_mul_pd(velec,rinvsq12);
 664
 665             /* Update potential sum for this i atom from the interaction with this j atom. */
 666             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 667             velecsum         = _mm_add_pd(velecsum,velec);
 668
 669             fscal            = felec;
 670
 671             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 672
 673             /* Update vectorial force */
 674             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 675             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 676             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 677
 678             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 679             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 680             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 681
 682             /**************************
 683              * CALCULATE INTERACTIONS *
 684              **************************/
 685
 686             /* COULOMB ELECTROSTATICS */
 687             velec            = _mm_mul_pd(qq20,rinv20);
 688             felec            = _mm_mul_pd(velec,rinvsq20);
 689
 690             /* Update potential sum for this i atom from the interaction with this j atom. */
 691             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 692             velecsum         = _mm_add_pd(velecsum,velec);
 693
 694             fscal            = felec;
 695
 696             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 697
 698             /* Update vectorial force */
 699             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 700             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 701             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 702
 703             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 704             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 705             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 706
 707             /**************************
 708              * CALCULATE INTERACTIONS *
 709              **************************/
 710
 711             /* COULOMB ELECTROSTATICS */
 712             velec            = _mm_mul_pd(qq21,rinv21);
 713             felec            = _mm_mul_pd(velec,rinvsq21);
 714
 715             /* Update potential sum for this i atom from the interaction with this j atom. */
 716             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 717             velecsum         = _mm_add_pd(velecsum,velec);
 718
 719             fscal            = felec;
 720
 721             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 722
 723             /* Update vectorial force */
 724             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 725             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 726             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 727
 728             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 729             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 730             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 731
 732             /**************************
 733              * CALCULATE INTERACTIONS *
 734              **************************/
 735
 736             /* COULOMB ELECTROSTATICS */
 737             velec            = _mm_mul_pd(qq22,rinv22);
 738             felec            = _mm_mul_pd(velec,rinvsq22);
 739
 740             /* Update potential sum for this i atom from the interaction with this j atom. */
 741             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 742             velecsum         = _mm_add_pd(velecsum,velec);
 743
 744             fscal            = felec;
 745
 746             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 747
 748             /* Update vectorial force */
 749             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 750             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 751             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 752
 753             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 754             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 755             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 756
 757             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 758
 759             /* Inner loop uses 279 flops */
 760         }
 761
 762         /* End of innermost loop */
 763
 764         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 765                                               f+i_coord_offset,fshift+i_shift_offset);
 766
 767         ggid                        = gid[iidx];
 768         /* Update potential energies */
 769         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 770
 771         /* Increment number of inner iterations */
 772         inneriter                  += j_index_end - j_index_start;
 773
 774         /* Outer loop uses 19 flops */
 775     }
 776
 777     /* Increment number of outer iterations */
 778     outeriter        += nri;
 779
 780     /* Update outer/inner flops */
 781
 782     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_VF,outeriter*19 + inneriter*279);
 783 }
 784 /*
 785  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW3W3_F_avx_128_fma_double
 786  * Electrostatics interaction: Coulomb
 787  * VdW interaction:            None
 788  * Geometry:                   Water3-Water3
 789  * Calculate force/pot:        Force
 790  */
 791 void
 792 nb_kernel_ElecCoul_VdwNone_GeomW3W3_F_avx_128_fma_double
 793                     (t_nblist * gmx_restrict                nlist,
 794                      rvec * gmx_restrict                    xx,
 795                      rvec * gmx_restrict                    ff,
 796                      t_forcerec * gmx_restrict              fr,
 797                      t_mdatoms * gmx_restrict               mdatoms,
 798                      nb_kernel_data_t * gmx_restrict        kernel_data,
 799                      t_nrnb * gmx_restrict                  nrnb)
 800 {
 801     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 802      * just 0 for non-waters.
 803      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 804      * jnr indices corresponding to data put in the four positions in the SIMD register.
 805      */
 806     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 807     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 808     int              jnrA,jnrB;
 809     int              j_coord_offsetA,j_coord_offsetB;
 810     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 811     real             rcutoff_scalar;
 812     real             *shiftvec,*fshift,*x,*f;
 813     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 814     int              vdwioffset0;
 815     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 816     int              vdwioffset1;
 817     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 818     int              vdwioffset2;
 819     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 820     int              vdwjidx0A,vdwjidx0B;
 821     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 822     int              vdwjidx1A,vdwjidx1B;
 823     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 824     int              vdwjidx2A,vdwjidx2B;
 825     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 826     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 827     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
 828     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 829     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 830     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 831     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 832     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 833     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 834     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 835     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 836     real             *charge;
 837     __m128d          dummy_mask,cutoff_mask;
 838     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 839     __m128d          one     = _mm_set1_pd(1.0);
 840     __m128d          two     = _mm_set1_pd(2.0);
 841     x                = xx[0];
 842     f                = ff[0];
 843
 844     nri              = nlist->nri;
 845     iinr             = nlist->iinr;
 846     jindex           = nlist->jindex;
 847     jjnr             = nlist->jjnr;
 848     shiftidx         = nlist->shift;
 849     gid              = nlist->gid;
 850     shiftvec         = fr->shift_vec[0];
 851     fshift           = fr->fshift[0];
 852     facel            = _mm_set1_pd(fr->epsfac);
 853     charge           = mdatoms->chargeA;
 854
 855     /* Setup water-specific parameters */
 856     inr              = nlist->iinr[0];
 857     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 858     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 859     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 860
 861     jq0              = _mm_set1_pd(charge[inr+0]);
 862     jq1              = _mm_set1_pd(charge[inr+1]);
 863     jq2              = _mm_set1_pd(charge[inr+2]);
 864     qq00             = _mm_mul_pd(iq0,jq0);
 865     qq01             = _mm_mul_pd(iq0,jq1);
 866     qq02             = _mm_mul_pd(iq0,jq2);
 867     qq10             = _mm_mul_pd(iq1,jq0);
 868     qq11             = _mm_mul_pd(iq1,jq1);
 869     qq12             = _mm_mul_pd(iq1,jq2);
 870     qq20             = _mm_mul_pd(iq2,jq0);
 871     qq21             = _mm_mul_pd(iq2,jq1);
 872     qq22             = _mm_mul_pd(iq2,jq2);
 873
 874     /* Avoid stupid compiler warnings */
 875     jnrA = jnrB = 0;
 876     j_coord_offsetA = 0;
 877     j_coord_offsetB = 0;
 878
 879     outeriter        = 0;
 880     inneriter        = 0;
 881
 882     /* Start outer loop over neighborlists */
 883     for(iidx=0; iidx<nri; iidx++)
 884     {
 885         /* Load shift vector for this list */
 886         i_shift_offset   = DIM*shiftidx[iidx];
 887
 888         /* Load limits for loop over neighbors */
 889         j_index_start    = jindex[iidx];
 890         j_index_end      = jindex[iidx+1];
 891
 892         /* Get outer coordinate index */
 893         inr              = iinr[iidx];
 894         i_coord_offset   = DIM*inr;
 895
 896         /* Load i particle coords and add shift vector */
 897         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 898                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 899
 900         fix0             = _mm_setzero_pd();
 901         fiy0             = _mm_setzero_pd();
 902         fiz0             = _mm_setzero_pd();
 903         fix1             = _mm_setzero_pd();
 904         fiy1             = _mm_setzero_pd();
 905         fiz1             = _mm_setzero_pd();
 906         fix2             = _mm_setzero_pd();
 907         fiy2             = _mm_setzero_pd();
 908         fiz2             = _mm_setzero_pd();
 909
 910         /* Start inner kernel loop */
 911         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 912         {
 913
 914             /* Get j neighbor index, and coordinate index */
 915             jnrA             = jjnr[jidx];
 916             jnrB             = jjnr[jidx+1];
 917             j_coord_offsetA  = DIM*jnrA;
 918             j_coord_offsetB  = DIM*jnrB;
 919
 920             /* load j atom coordinates */
 921             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 922                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 923
 924             /* Calculate displacement vector */
 925             dx00             = _mm_sub_pd(ix0,jx0);
 926             dy00             = _mm_sub_pd(iy0,jy0);
 927             dz00             = _mm_sub_pd(iz0,jz0);
 928             dx01             = _mm_sub_pd(ix0,jx1);
 929             dy01             = _mm_sub_pd(iy0,jy1);
 930             dz01             = _mm_sub_pd(iz0,jz1);
 931             dx02             = _mm_sub_pd(ix0,jx2);
 932             dy02             = _mm_sub_pd(iy0,jy2);
 933             dz02             = _mm_sub_pd(iz0,jz2);
 934             dx10             = _mm_sub_pd(ix1,jx0);
 935             dy10             = _mm_sub_pd(iy1,jy0);
 936             dz10             = _mm_sub_pd(iz1,jz0);
 937             dx11             = _mm_sub_pd(ix1,jx1);
 938             dy11             = _mm_sub_pd(iy1,jy1);
 939             dz11             = _mm_sub_pd(iz1,jz1);
 940             dx12             = _mm_sub_pd(ix1,jx2);
 941             dy12             = _mm_sub_pd(iy1,jy2);
 942             dz12             = _mm_sub_pd(iz1,jz2);
 943             dx20             = _mm_sub_pd(ix2,jx0);
 944             dy20             = _mm_sub_pd(iy2,jy0);
 945             dz20             = _mm_sub_pd(iz2,jz0);
 946             dx21             = _mm_sub_pd(ix2,jx1);
 947             dy21             = _mm_sub_pd(iy2,jy1);
 948             dz21             = _mm_sub_pd(iz2,jz1);
 949             dx22             = _mm_sub_pd(ix2,jx2);
 950             dy22             = _mm_sub_pd(iy2,jy2);
 951             dz22             = _mm_sub_pd(iz2,jz2);
 952
 953             /* Calculate squared distance and things based on it */
 954             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 955             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 956             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 957             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 958             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 959             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 960             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 961             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 962             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 963
 964             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 965             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 966             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 967             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 968             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 969             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 970             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 971             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 972             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 973
 974             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 975             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 976             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 977             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 978             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 979             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 980             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 981             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 982             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 983
 984             fjx0             = _mm_setzero_pd();
 985             fjy0             = _mm_setzero_pd();
 986             fjz0             = _mm_setzero_pd();
 987             fjx1             = _mm_setzero_pd();
 988             fjy1             = _mm_setzero_pd();
 989             fjz1             = _mm_setzero_pd();
 990             fjx2             = _mm_setzero_pd();
 991             fjy2             = _mm_setzero_pd();
 992             fjz2             = _mm_setzero_pd();
 993
 994             /**************************
 995              * CALCULATE INTERACTIONS *
 996              **************************/
 997
 998             /* COULOMB ELECTROSTATICS */
 999             velec            = _mm_mul_pd(qq00,rinv00);
1000             felec            = _mm_mul_pd(velec,rinvsq00);
1001
1002             fscal            = felec;
1003
1004             /* Update vectorial force */
1005             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1006             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1007             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1008
1009             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1010             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1011             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1012
1013             /**************************
1014              * CALCULATE INTERACTIONS *
1015              **************************/
1016
1017             /* COULOMB ELECTROSTATICS */
1018             velec            = _mm_mul_pd(qq01,rinv01);
1019             felec            = _mm_mul_pd(velec,rinvsq01);
1020
1021             fscal            = felec;
1022
1023             /* Update vectorial force */
1024             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1025             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1026             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1027
1028             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1029             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1030             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1031
1032             /**************************
1033              * CALCULATE INTERACTIONS *
1034              **************************/
1035
1036             /* COULOMB ELECTROSTATICS */
1037             velec            = _mm_mul_pd(qq02,rinv02);
1038             felec            = _mm_mul_pd(velec,rinvsq02);
1039
1040             fscal            = felec;
1041
1042             /* Update vectorial force */
1043             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1044             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1045             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1046
1047             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1048             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1049             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1050
1051             /**************************
1052              * CALCULATE INTERACTIONS *
1053              **************************/
1054
1055             /* COULOMB ELECTROSTATICS */
1056             velec            = _mm_mul_pd(qq10,rinv10);
1057             felec            = _mm_mul_pd(velec,rinvsq10);
1058
1059             fscal            = felec;
1060
1061             /* Update vectorial force */
1062             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1063             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1064             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1065
1066             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1067             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1068             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1069
1070             /**************************
1071              * CALCULATE INTERACTIONS *
1072              **************************/
1073
1074             /* COULOMB ELECTROSTATICS */
1075             velec            = _mm_mul_pd(qq11,rinv11);
1076             felec            = _mm_mul_pd(velec,rinvsq11);
1077
1078             fscal            = felec;
1079
1080             /* Update vectorial force */
1081             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1082             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1083             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1084
1085             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1086             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1087             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1088
1089             /**************************
1090              * CALCULATE INTERACTIONS *
1091              **************************/
1092
1093             /* COULOMB ELECTROSTATICS */
1094             velec            = _mm_mul_pd(qq12,rinv12);
1095             felec            = _mm_mul_pd(velec,rinvsq12);
1096
1097             fscal            = felec;
1098
1099             /* Update vectorial force */
1100             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1101             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1102             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1103
1104             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1105             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1106             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1107
1108             /**************************
1109              * CALCULATE INTERACTIONS *
1110              **************************/
1111
1112             /* COULOMB ELECTROSTATICS */
1113             velec            = _mm_mul_pd(qq20,rinv20);
1114             felec            = _mm_mul_pd(velec,rinvsq20);
1115
1116             fscal            = felec;
1117
1118             /* Update vectorial force */
1119             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1120             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1121             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1122
1123             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1124             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1125             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1126
1127             /**************************
1128              * CALCULATE INTERACTIONS *
1129              **************************/
1130
1131             /* COULOMB ELECTROSTATICS */
1132             velec            = _mm_mul_pd(qq21,rinv21);
1133             felec            = _mm_mul_pd(velec,rinvsq21);
1134
1135             fscal            = felec;
1136
1137             /* Update vectorial force */
1138             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1139             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1140             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1141
1142             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1143             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1144             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1145
1146             /**************************
1147              * CALCULATE INTERACTIONS *
1148              **************************/
1149
1150             /* COULOMB ELECTROSTATICS */
1151             velec            = _mm_mul_pd(qq22,rinv22);
1152             felec            = _mm_mul_pd(velec,rinvsq22);
1153
1154             fscal            = felec;
1155
1156             /* Update vectorial force */
1157             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1158             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1159             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1160
1161             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1162             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1163             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1164
1165             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1166
1167             /* Inner loop uses 270 flops */
1168         }
1169
1170         if(jidx<j_index_end)
1171         {
1172
1173             jnrA             = jjnr[jidx];
1174             j_coord_offsetA  = DIM*jnrA;
1175
1176             /* load j atom coordinates */
1177             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1178                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1179
1180             /* Calculate displacement vector */
1181             dx00             = _mm_sub_pd(ix0,jx0);
1182             dy00             = _mm_sub_pd(iy0,jy0);
1183             dz00             = _mm_sub_pd(iz0,jz0);
1184             dx01             = _mm_sub_pd(ix0,jx1);
1185             dy01             = _mm_sub_pd(iy0,jy1);
1186             dz01             = _mm_sub_pd(iz0,jz1);
1187             dx02             = _mm_sub_pd(ix0,jx2);
1188             dy02             = _mm_sub_pd(iy0,jy2);
1189             dz02             = _mm_sub_pd(iz0,jz2);
1190             dx10             = _mm_sub_pd(ix1,jx0);
1191             dy10             = _mm_sub_pd(iy1,jy0);
1192             dz10             = _mm_sub_pd(iz1,jz0);
1193             dx11             = _mm_sub_pd(ix1,jx1);
1194             dy11             = _mm_sub_pd(iy1,jy1);
1195             dz11             = _mm_sub_pd(iz1,jz1);
1196             dx12             = _mm_sub_pd(ix1,jx2);
1197             dy12             = _mm_sub_pd(iy1,jy2);
1198             dz12             = _mm_sub_pd(iz1,jz2);
1199             dx20             = _mm_sub_pd(ix2,jx0);
1200             dy20             = _mm_sub_pd(iy2,jy0);
1201             dz20             = _mm_sub_pd(iz2,jz0);
1202             dx21             = _mm_sub_pd(ix2,jx1);
1203             dy21             = _mm_sub_pd(iy2,jy1);
1204             dz21             = _mm_sub_pd(iz2,jz1);
1205             dx22             = _mm_sub_pd(ix2,jx2);
1206             dy22             = _mm_sub_pd(iy2,jy2);
1207             dz22             = _mm_sub_pd(iz2,jz2);
1208
1209             /* Calculate squared distance and things based on it */
1210             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1211             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1212             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1213             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1214             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1215             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1216             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1217             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1218             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1219
1220             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1221             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1222             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1223             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1224             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1225             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1226             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1227             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1228             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1229
1230             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1231             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1232             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1233             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1234             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1235             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1236             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1237             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1238             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1239
1240             fjx0             = _mm_setzero_pd();
1241             fjy0             = _mm_setzero_pd();
1242             fjz0             = _mm_setzero_pd();
1243             fjx1             = _mm_setzero_pd();
1244             fjy1             = _mm_setzero_pd();
1245             fjz1             = _mm_setzero_pd();
1246             fjx2             = _mm_setzero_pd();
1247             fjy2             = _mm_setzero_pd();
1248             fjz2             = _mm_setzero_pd();
1249
1250             /**************************
1251              * CALCULATE INTERACTIONS *
1252              **************************/
1253
1254             /* COULOMB ELECTROSTATICS */
1255             velec            = _mm_mul_pd(qq00,rinv00);
1256             felec            = _mm_mul_pd(velec,rinvsq00);
1257
1258             fscal            = felec;
1259
1260             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1261
1262             /* Update vectorial force */
1263             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1264             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1265             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1266
1267             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1268             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1269             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1270
1271             /**************************
1272              * CALCULATE INTERACTIONS *
1273              **************************/
1274
1275             /* COULOMB ELECTROSTATICS */
1276             velec            = _mm_mul_pd(qq01,rinv01);
1277             felec            = _mm_mul_pd(velec,rinvsq01);
1278
1279             fscal            = felec;
1280
1281             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1282
1283             /* Update vectorial force */
1284             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1285             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1286             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1287
1288             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1289             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1290             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1291
1292             /**************************
1293              * CALCULATE INTERACTIONS *
1294              **************************/
1295
1296             /* COULOMB ELECTROSTATICS */
1297             velec            = _mm_mul_pd(qq02,rinv02);
1298             felec            = _mm_mul_pd(velec,rinvsq02);
1299
1300             fscal            = felec;
1301
1302             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1303
1304             /* Update vectorial force */
1305             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1306             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1307             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1308
1309             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1310             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1311             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1312
1313             /**************************
1314              * CALCULATE INTERACTIONS *
1315              **************************/
1316
1317             /* COULOMB ELECTROSTATICS */
1318             velec            = _mm_mul_pd(qq10,rinv10);
1319             felec            = _mm_mul_pd(velec,rinvsq10);
1320
1321             fscal            = felec;
1322
1323             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1324
1325             /* Update vectorial force */
1326             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1327             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1328             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1329
1330             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1331             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1332             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1333
1334             /**************************
1335              * CALCULATE INTERACTIONS *
1336              **************************/
1337
1338             /* COULOMB ELECTROSTATICS */
1339             velec            = _mm_mul_pd(qq11,rinv11);
1340             felec            = _mm_mul_pd(velec,rinvsq11);
1341
1342             fscal            = felec;
1343
1344             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1345
1346             /* Update vectorial force */
1347             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1348             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1349             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1350
1351             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1352             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1353             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1354
1355             /**************************
1356              * CALCULATE INTERACTIONS *
1357              **************************/
1358
1359             /* COULOMB ELECTROSTATICS */
1360             velec            = _mm_mul_pd(qq12,rinv12);
1361             felec            = _mm_mul_pd(velec,rinvsq12);
1362
1363             fscal            = felec;
1364
1365             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1366
1367             /* Update vectorial force */
1368             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1369             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1370             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1371
1372             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1373             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1374             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1375
1376             /**************************
1377              * CALCULATE INTERACTIONS *
1378              **************************/
1379
1380             /* COULOMB ELECTROSTATICS */
1381             velec            = _mm_mul_pd(qq20,rinv20);
1382             felec            = _mm_mul_pd(velec,rinvsq20);
1383
1384             fscal            = felec;
1385
1386             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1387
1388             /* Update vectorial force */
1389             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1390             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1391             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1392
1393             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1394             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1395             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1396
1397             /**************************
1398              * CALCULATE INTERACTIONS *
1399              **************************/
1400
1401             /* COULOMB ELECTROSTATICS */
1402             velec            = _mm_mul_pd(qq21,rinv21);
1403             felec            = _mm_mul_pd(velec,rinvsq21);
1404
1405             fscal            = felec;
1406
1407             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1408
1409             /* Update vectorial force */
1410             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1411             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1412             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1413
1414             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1415             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1416             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1417
1418             /**************************
1419              * CALCULATE INTERACTIONS *
1420              **************************/
1421
1422             /* COULOMB ELECTROSTATICS */
1423             velec            = _mm_mul_pd(qq22,rinv22);
1424             felec            = _mm_mul_pd(velec,rinvsq22);
1425
1426             fscal            = felec;
1427
1428             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1429
1430             /* Update vectorial force */
1431             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1432             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1433             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1434
1435             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1436             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1437             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1438
1439             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1440
1441             /* Inner loop uses 270 flops */
1442         }
1443
1444         /* End of innermost loop */
1445
1446         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1447                                               f+i_coord_offset,fshift+i_shift_offset);
1448
1449         /* Increment number of inner iterations */
1450         inneriter                  += j_index_end - j_index_start;
1451
1452         /* Outer loop uses 18 flops */
1453     }
1454
1455     /* Increment number of outer iterations */
1456     outeriter        += nri;
1457
1458     /* Update outer/inner flops */
1459
1460     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_F,outeriter*18 + inneriter*270);
1461 }